Overview: Bland-Altman Method in EEG Device & Telemetry Validation
In biomedical signal processing and clinical electroencephalography (EEG), evaluating the clinical interchangeability and measurement agreement of two acquisition hardware devices, recording telemetry systems, or electrode montage channels is a foundational task. While traditional bivariate correlation coefficients such as Pearson's $r$ or Spearman's $\rho$ measure the strength of a linear association between two measurement sets $S_1$ and $S_2$, they fail to detect systematic bias, constant calibration offsets, or proportional magnitude-dependent variances. Two EEG amplifiers can exhibit a near-perfect correlation coefficient ($r > 0.99$) while one systematically overestimates scalp microvolt potentials by $30\%$ or introduces a constant offset of $15\text{ }\mu\text{V}$.
To resolve this statistical limitation, Douglas Bland and Richard Altman proposed the Bland-Altman plot (also known as the Tukey mean-difference plot). In electroencephalography, this graphical technique maps the mean measurement value of paired observations along the horizontal axis ($X_i = \frac{S_{1,i} + S_{2,i}}{2}$) against their arithmetic difference along the vertical axis ($Y_i = S_{1,i} - S_{2,i}$). By decoupling the magnitude of the neurophysiological activity from the discrepancy between measurement techniques, the Bland-Altman method reveals systematic mean bias ($\bar{d}$), the random dispersion of differences ($s_d$), and the limits of agreement ($\text{LOA} = \bar{d} \pm 1.96 s_d$).
When applied to EEG signal streams, Bland-Altman analysis allows neuroscientists and biomedical engineers to quantify whether novel dry-sensor EEG headsets agree with clinical wet $10\text{--}20$ gold-cup electrodes, assess inter-channel stability across cerebral hemispheres (such as $F_3$ versus $F_4$), and verify if algorithmic artifact clearance pipelines preserve underlying spectral band power (alpha, beta, theta, and delta rhythms) without introducing spectral amplitude distortion.
Technical & Mathematical Foundations
The primary mathematical formulation of the Bland-Altman agreement model assumes paired scalar measurements $S_{1,i}$ and $S_{2,i}$ sampled across $N$ continuous time points or discrete epoch segments.
For each observation $i \in \{1, 2, \dots, N\}$, the mean value $A_i$ and the difference $D_i$ are computed as:
$$A_i = \frac{S_{1,i} + S_{2,i}}{2}, \quad D_i = S_{1,i} - S_{2,i}$$
The estimated systematic bias ($\bar{d}$) representing constant calibration discrepancy is defined by the sample arithmetic mean of differences:
$$\bar{d} = \frac{1}{N} \sum_{i=1}^{N} D_i$$
The sample standard deviation of differences ($s_d$) quantifies random measurement variability:
$$s_d = \sqrt{\frac{1}{N - 1} \sum_{i=1}^{N} \left(D_i - \bar{d}\right)^2}$$
Under the assumption of normally distributed differences ($D_i \sim \mathcal{N}(\bar{d}, s_d^2)$), approximately $95\%$ of all measurement differences fall within the parametric Limits of Agreement (LOA):
$$\text{LOA}_{\text{upper}} = \bar{d} + 1.96 \cdot s_d, \quad \text{LOA}_{\text{lower}} = \bar{d} - 1.96 \cdot s_d$$
To quantify the statistical precision of these estimated parameters, $95\%$ confidence intervals (CI) are calculated using the standard error of the mean bias $SE(\bar{d})$ and standard error of the limits of agreement $SE(\text{LOA})$:
$$SE(\bar{d}) = \frac{s_d}{\sqrt{N}}, \quad SE(\text{LOA}) \approx \sqrt{\frac{3 \cdot s_d^2}{N}} \approx 1.71 \cdot \frac{s_d}{\sqrt{N}}$$
$$\text{CI}_{95\%}(\bar{d}) = \bar{d} \pm t_{N-1, 0.025} \cdot SE(\bar{d})$$
$$\text{CI}_{95\%}(\text{LOA}) = \text{LOA} \pm t_{N-1, 0.025} \cdot SE(\text{LOA})$$
When proportional bias exists (for instance, when amplifier saturation or nonlinear gain causes difference magnitude to increase with EEG voltage), ordinary least squares (OLS) linear regression models $D_i$ against $A_i$:
$$D_i = \beta_0 + \beta_1 A_i + \varepsilon_i$$
Where $\beta_1$ represents the proportional bias slope coefficient:
$$\beta_1 = \frac{\sum_{i=1}^N (A_i - \bar{A})(D_i - \bar{D})}{\sum_{i=1}^N (A_i - \bar{A})^2}$$
When differences exhibit extreme skewness or non-Gaussian distribution due to transient EEG artifacts (such as high-voltage ocular blinks or cardiac $R$-wave breakthrough), non-parametric limits of agreement are computed directly from the $2.5^{\text{th}}$ and $97.5^{\text{th}}$ empirical percentiles of the sorted difference vector $D_{(i)}$. Furthermore, for non-linear power metrics where errors are multiplicative rather than additive, the logarithmic ratio transformation $D_i^{\log} = \ln(S_{1,i}) - \ln(S_{2,i}) = \ln\left(\frac{S_{1,i}}{S_{2,i}}\right)$ is computed, allowing agreement limits to be exponentiated back into percentage differences ($e^{\bar{d}_{\log} \pm 1.96 s_{\log}} - 1$).